Leveraging BERT for embedding ICD codes from large scale cardiovascular EMR data to understand patient diagnostic...

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Informatics
• Cardiovascular Research

Background:

• Electronic medical records (EMRs) and AI integration are crucial for real-world evidence (RWE) studies.
• Patient characterization relies on extracting insights from coded medical data like ICD-10 codes.
• Traditional methods like one-hot encoding (OHE) struggle with high-dimensional medical data.

Purpose of the Study:

• To introduce and evaluate a Bidirectional Encoder Representations from Transformers (BERT) approach for encoding ICD-10 diagnostic codes.
• To compare the performance of ClinicalBERT embeddings against OHE for patient characterization and outcome prediction.
• To assess the impact of AI-driven code embedding on dimensionality reduction and predictive accuracy in cardiology.

Main Methods:

• Utilized a Bidirectional Encoder Representations from Transformers (BERT) model, specifically ClinicalBERT, to generate embeddings for ICD-10 codes.
• Compared ClinicalBERT embeddings with traditional one-hot encoding (OHE) using data from 495,269 cardiology patients (2000-2020).
• Evaluated model performance in predicting major adverse cardiovascular events within one year post-PCI or CABG.

Main Results:

• The ClinicalBERT (code-embedded) model significantly outperformed OHE in predicting major adverse cardiovascular events (AUC 0.746 vs. 0.719).
• ClinicalBERT embeddings drastically reduced data dimensionality from 2,492 to 128 features.
• The integrated approach using diagnostic and medication data enhanced predictive precision.

Conclusions:

• AI-driven embedding of ICD-10 codes using ClinicalBERT offers superior performance and efficiency over traditional methods for patient characterization.
• This approach enhances the precision of predicting adverse cardiovascular events, aiding clinical decision-making.
• The integration of AI with EMRs, particularly through advanced natural language processing techniques, holds significant promise for advancing medical research and patient care.